Discrete Mathematics 342 (2019) 988–996

Contents lists available at ScienceDirect

Discrete Mathematics
journal homepage: www.elsevier.com/locate/disc

H-free subgraphs of dense graphs maximizing the number of

cliques and their blow-ups
∗
Noga Alon a ,1 , Clara Shikhelman b ,2 ,
a
Sackler School of Mathematics and Blavatnik School of Computer Science, Tel Aviv University, Tel Aviv 69978, Israel
b
Sackler School of Mathematics, Tel Aviv University, Tel Aviv 69978, Israel

article info a b s t r a c t

Article history: We consider the structure of H-free subgraphs of graphs with high minimal degree. We
Received 20 June 2017 prove that for every k > m there exists an ϵ := ϵ (k, m) > 0 so that the following holds. For
Received in revised form 18 September 2018 every graph H with chromatic number k from which one can delete an edge and reduce the
Accepted 15 November 2018
chromatic number, and for every graph G on n > n0 (H) vertices in which all degrees are
Available online 24 December 2018
at least (1 − ϵ )n, any subgraph of G which is H-free and contains the maximum number of
Keywords: copies of the complete graph Km is (k − 1)-colorable.
Turan type problems We also consider several extensions for the case of a general forbidden graph H of a
Chromatic number given chromatic number, and for subgraphs maximizing the number of copies of balanced
Dense graphs blowups of complete graphs.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction

The well known theorem of Turán [18] states that a Kk -free subgraph of the complete graph on n vertices with the
maximum possible number of edges is (k − 1)-chromatic. Erdős, Stone and Simonovits show in [13,11] that for general
H with χ (H) = k the maximum possible number of edges in an H-free graph on n vertices is at most o(n2 ) more than the
number of edges in a (k − 1)-chromatic graph on n vertices. In [4] it is shown that the same holds for H-free subgraphs of
the complete graph that have the maximum possible number of copies of Km for a fixed m such that k > m ≥ 2.
Looking at subgraphs of general graphs G it is clear that a Km -free subgraph of G with the maximum possible number of
edges has at least as many edges as the largest (m − 1)-partite subgraph. In [10] Erdős asked for which graphs there is an
equality between the two. In [1] it is shown that this is the case for line graphs of bipartite graphs. In a different direction,
in [8] it is proved that if a graph has a high enough minimum degree then any subgraph of it which is K3 -free and has the
maximum possible number of edges is bipartite. In [7] a stronger bound is given on the minimum degree ensuring this.
Before stating a generalization of these theorems we introduce some notation.
For a graph G, fixed graphs H and T and an integer k let Gpart(k),T be a k-partite subgraph of G with the maximum possible
number of copies of T and let Gex (T , H) be the family of subgraphs of G that are H-free and have the maximum possible
number of copies of T . Let N (G, T ) denote the number of copies of T in G. Call a graph H edge critical if there is an edge {u, v}
in E(H) whose removal reduces the chromatic number of H.
In [3] the following theorem is proved, generalizing the results in [8] and [7]. Throughout the paper we denote by δ (G)
the minimum degree in the graph G.

∗ Corresponding author.
E-mail addresses: nogaa@tau.ac.il (N. Alon), clarashk@post.tau.ac.il (C. Shikhelman).
1 Research supported in part by a BSF grant, an ISF grant and a GIF grant.
2 Research supported in part by an ISF grant.

https://doi.org/10.1016/j.disc.2018.11.012
0012-365X/© 2018 Elsevier B.V. All rights reserved.
 N. Alon and C. Shikhelman / Discrete Mathematics 342 (2019) 988–996 989

Theorem 1.1 ([3]). Let H be a graph with χ (H) = k + 1. Then there are positive constants γ := γ (H) and µ := µ(H) such that
if G is a graph on n > n0 (H) vertices with δ (G) > (1 − µ)n then for every Gex ∈ Gex (K2 , H)

1. If H is edge critical then N (Gpart(k),K2 , K2 ) = N (Gex , K2 ).

2. Otherwise, N (Gpart(k),K2 , K2 ) ≤ N (Gex , K2 ) ≤ N (Gpart(k),K2 , K2 ) + O(n2−γ ).
In the present short paper we prove two theorems for H-free subgraphs assuming H is edge critical. The first is for
subgraphs maximizing the number of copies of Km and the second for subgraphs maximizing the number of blow-ups of
Km . We also establish a proposition concerning graphs H that are not edge critical.

Theorem 1.2. For every two integers k > m and every edge critical graph H such that χ (H) = k there exist constants
ϵ := ϵ (k, m) > 0 and n0 = n0 (H) such that the following holds. Let G be a graph on n > n0 vertices with δ (G) ≥ (1 − ϵ )n, then
for every Gex ∈ Gex (Km , H) the graph Gex is (k − 1)-colorable.
For integers m and t let Km (t) denote the t-blow-up of Km , that is, the graph obtained by replacing each vertex of Km by
an independent set of size t and each edge by a complete bipartite graph between the corresponding independent sets.

Theorem 1.3. For integers m and t and every edge critical H such that χ (H) = m + 1 there exist constants ϵ := ϵ (m, t) and
n0 := n0 (H) such that the following holds. Let G be a graph on n > n0 vertices with δ (G) > (1 −ϵ )n, then every Gex ∈ Gex (Km (t), H)
is m-colorable.
Finally, for graphs H which are not edge critical we prove the following.

Proposition 1.4. For every integers m < k and t and graph H such that χ (H) = k there exists ϵ := ϵ (m, t , k) and n0 := n0 (H)
such that the following holds. Let G be a graph on n > n0 vertices with δ (G) > (1 − ϵ )n and assume that t = 1 or k = m + 1,
then every Gex ∈ Gex (Km (t), H) can be made (k − 1)-colorable by deleting o(n2 ) edges.
Theorems 1.2 and 1.3 cannot be directly generalized to graphs H that are not edge critical as we can add to any (k − 1)-
partite graph an edge without creating a copy of such H. On the other hand, we believe that the error term o(n2 ) in
Proposition 1.4 can be improved to O(n2−δ ) for some δ := δ (H).
The rest of this short paper is organized as follows. In Section 2 we state several known results and prove some helpful
lemmas. Section 3 contains the proof of Theorem 1.2. Theorem 1.3 is proved in Section 4 and the proof of Proposition 1.4
appears in Section 5. Section 6 contains some concluding remarks and open problems.

2. Preliminary results

We start by stating several results about H-free graphs with high degrees and by deducing a corollary. Some of the
theorems stated are simplified versions of the original results.
The first result about Kk -free graphs is by Andrásfai, Erdős and Sós.

Theorem 2.1 ([6]). Let G be a graph on n vertices. If G is Kk -free and δ (G) ≥ 1 − 3

n then χ (G) ≤ k − 1.
( )
3k−4

A generalization of Theorem 2.1 proved in [12] is the following.

Theorem 2.2 ([12]). Let H be a fixed edge critical graph which ( is not Kk )and assume χ (H) = k. If G is a graph on n > n0 (H)
vertices which is H-free and contains a copy of Kk then δ (G) ≤ 1 − k−13/2 n + O(1).

This implies that if n is large enough, δ (G) ≥ 1 − 3k3−4 n ≥ 1 − k−13/2 n + O(1), and if G is H-free for some edge critical
( ) ( )
graph H with χ (H) = k then it must also be Kk -free. Together with Theorem 2.1 we get the following corollary:

) edge critical graph such that χ (H) = k. Let G be a graph on n > n0 (H) vertices which is H-free
Corollary 2.3. Let H( be a fixed
and satisfies δ (G) ≥ 1 − 3k3−4 n, then χ (G) ≤ k − 1.
We next state the graph removal lemma as it appears in [9] (see also [2,17] and [14]) and prove a simple lemma using it.
Throughout the paper we denote by v (G) the number of vertices in the graph G.

Theorem 2.4 (The Graph Removal Lemma). For any graph H with v (H) vertices and any ϵ > 0, there exists a δ > 0 such that any
graph on n vertices which contains at most δ nv (H) copies of H can be made H-free by removing at most ϵ n2 edges.
Throughout the paper, for fixed graphs T and H and an integer n we denote by ex(n, T , H) the maximum possible number
of copies of T in an H-free graph on n vertices. In [4] this function is studied for various T and H. We will use the following
result and a simple corollary of it.

Proposition 2.5 ([4]). Let T be a fixed graph with t vertices. Then ex(n, T , H) = Ω (nt ) if and only if H is not a subgraph of a
blow-up of T . Otherwise, ex(n, T , H) ≤ nt −ϵ (T ,H) for some ϵ (T , H) > 0.
990 N. Alon and C. Shikhelman / Discrete Mathematics 342 (2019) 988–996

Corollary 2.6. Let H be a fixed graph such that χ (H) = k and let G be an H-free graph on n vertices, where n > n0 (H). Then G
can be made Kk -free by deleting o(n2 ) edges.

This holds as H is contained in a blow-up of Kk and hence N (G, Kk ) ≤ ex(n, Kk , H) ≤ o(nk ). By the graph removal lemma
G can be made Kk -free by removing o(n2 ) edges, as needed.
We next state and prove two additional technical lemmas.

Lemma 2.7. Let G be a graph on n vertices satisfying δ (G) > (1 − ϵ )n for some fixed 1 > ϵ > 0, and let m < k and t be integers.
m
1. N (Gpart(k−1),Km , Km ) ≥ 1 − c1′ ϵ .
( )(k−1)( n
)
m k−1
2. Let k = m + 1. Then N (Gpart(m),Km (t) , Km (t)) ≥ (1 + o(1))(1 − c2′ ϵ )nmt (t !m1t )m

where c1′ := c1′ (m) and c2′ := c2′ (m, t).

Proof. To prove part 1 note that as δ (G) ≥ (1 − ϵ )n the number of copies of Km in G is at least
1 ( )1
n · (1 − ϵ )n · (1 − 2ϵ )n . . . (1 − (m − 1)ϵ )n ≥ nm 1 − c1′ ϵ .
m! m!
Randomly partitioning the graph G into k − 1 sets yields a graph in which the expected number of copies of Km is at least:
) 1 k−2 k−3 k−m
nm 1 − c1′ ϵ ...
(
· ·
m! k − 1 k − 1 k−1
(k − 2) !
=nm 1 − c1′ ϵ
( )
m!(k − 1)m−1 (k − (m + 1))!
( )( )m
k −1 n
= 1 − c1′ ϵ .
( )
m k−1
Thus Gpart(k−1),Km should have at least that many copies. This proves Part 1.
Similarly, to prove part 2 observe that the number of copies of Km (t) in G is at least
1 nmt
m
(n)t ((1 − t ϵ )n)t · . . . · ((1 − (m − 1)t ϵ )n)t ≥ (1 + o(1)) (1 − c2′ ϵ ).
m!(t !) m!(t !)m
Randomly partitioning G into m parts gives a graph in which the expected number of copies of Km (t) is at least
nmt ( 1 )( m − 1 1 ) (1 1 )
(1 + o(1))(1 − c2′ ϵ ) ...
m!(t ! )m mt −1 m mt −1 m mt −1
1 m!
=(1 + o(1))(1 − c2′ ϵ )nmt
m!(t !)m mmt
1
=(1 + o(1))(1 − c2′ ϵ )n mt
(t !mt )m
and thus Gpart(m),Km must have at least that many copies of Km (t). □

For a graph G, a set of vertices U ⊂ V (G) and integers k, m and t, let fk,m,t (U) be the maximum number of copies of Km (t)
in a (k − 1)-partite subgraph of G[U ]

Lemma 2.8. Let G be a graph on n vertices with δ (G) > (1 − ϵ )n for some fixed ϵ > 0, let t and m < k be integers and let α > 0
be a fixed constant. Then there exist constants c1 := c1 (k, m, α ) and c2 := c2 (m, t , α ) such that for every set U ⊆ V (G) satisfying
|U | ≥ α n and every v ∈ V (G) \ U the following hold
− c1 ϵ )
(k−1)
1. fm,k,1 (U ∪ {v}) ≥ fm,k,1 (U) + (1 + o(1))|U |m−1 m (k−1)m
m
(1
2. fm,m+1,t (U ∪ {v}) ≥ fm,m+1,t (U) + (1 + o(1))|U | mt −1 mt
(t !mt )m
(1 − c2 ϵ )

Proof. Let |U | = q ≥ α n. We first prove part 1. Fix a partition of G[U ] into k − 1 parts with fm,k,1 (U) copies of Km . By
Lemma 2.7 part 1 and as δ (G[U ]) ≥ (1 − αϵ )q the number of copies of Km is at least:

ϵ) k−1
( )
( 1
(1 + o(1))qm 1 − c1′ .
α m (k − 1)m
Averaging we get that there is a vertex, say w ∈ U, so that the number of copies of Km it takes part in is at least
ϵ) k−1 ′ ϵ
( ) ) (k − 1) m
m ( 1 (
(1 + o(1)) qm 1 − c1′ = (1 + o(1))qm−1
1 − c .
q α m (k − 1)m 1
α m (k − 1)m
 N. Alon and C. Shikhelman / Discrete Mathematics 342 (2019) 988–996 991

Let U1 , . . . , Uk−1 be the above fixed partition of U which has fm,k,1 (U) copies of Km and assume, without loss of generality,
that w ∈ Uk−1 . We add v to Uk−1 and bound from below the number of copies of Km we ∑ add by doing this. ∑
Let bi = |Ui | and
let di be the number of neighbors w has in Ui which are not neighbors of v . Note that i∈[k−2] di ≤ ϵ n and i∈[k−2] bi ≤ q.
For each Ui we estimate the number of copies of Km in which w takes part that use vertices from Ui that are not neighbors
of v . There are di such vertices, and in the worst case each such vertex is connected to all of the sets Uj for j ̸ = i, k − 1. Thus
the number of copies of Km that w takes part in and v does not is at most
k−2
∑ ∑
bj1 . . . bjm−2
( )
di
i=1 {j1 ,...,jm−2 }⊆[k−2]\i

≤(d1 + · · · + dk−2 )(b1 + · · · + bk−2 )m−2

1
≤ϵ n · qm−2 ≤ ϵ qm−1 .
α
Thus when adding v the number of copies of Km added is at least

′ ϵ ϵ
( ) (k − 1) m
m−1
(1 + o(1))q [ 1 − c1 − ]
α m (k − 1)m α
( )
k−1 m
=(1 + o(1))qm−1 (1 − c1 ϵ ).
m (k − 1)m
The proof of Part 2 is similar. By Lemma 2.7 part 2, in any partition of G[U ] into m parts in which the number of copies of
Km (t) is fm,m+1,t (U), this number is at least
1
(1 + o(1)) (1 − c2′ ϵ )qmt .
(t !mt )m
Let U = U1 ∪ · · · ∪ Um be such a partition. By averaging there must be a vertex, say w ∈ U, such that the number of copies
of Km (t) it takes part in is at least:
mt 1 mt
(1 + o(1)) qmt (1 − c2′ ϵ ) = (1 + o(1))qmt −1 (1 − c2′ ϵ ).
q (t !mt )m (t !mt )m
Assume, without loss of generality, that w ∈ Um , and let us add v to Um . Let bi = |Ui | and let di be the number of vertices
in Ui that are neighbors of w and not of v . Then the number of copies of Km (t) in this partition that w takes part in and v does
not is at most
( m−1 (
)∑ ) ( )
bm bi ∏ bj
di
t −1 t −1 t
i=1 j∈[m−1]\i
m−1 ( )
∑ q
<( di ) < c2′′ ϵ qmt −1
mt − 2
i=1

where the last inequality is true for some c2′′ := c2′′ (m, t , α ). Thus when adding v to Um the number of copies of Km (t) added
is at least
mt mt
qmt −1 [ (1 − c2′ ϵ ) − c2′′ ϵ] = qmt −1 (1 − c2 ϵ )
(t !mt )m (t !mt )m
as needed. □

3. Maximizing the number of cliques

In the proof of Theorem 1.2 we use the following result from [4].

Proposition 3.1 ([4]). Let H be a graph such that χ (H) = k > m then
( )( )m
k−1 n
ex(n, Km , H) = (1 + o(1)) .
m k−1

Proof of Theorem 1.2. Let G and H be as in the theorem and let Gex ∈ Gex (Km , H). If δ (Gex ) ≥ (1 − 3k3−4 )n, as H is edge critical,
by Corollary 2.3 χ (Gex ) ≤ k − 1 and we are done. Thus assume towards contradiction that δ (Gex ) < (1 − 3k3−4 )n.
As any partition of G into k − 1 parts is H-free, by Lemma 2.7, part 1, the number of copies of Km in Gex must be at least
( )( )m
) k−1 n
1 − c1′ ϵ .
(
(1)
m k−1
992 N. Alon and C. Shikhelman / Discrete Mathematics 342 (2019) 988–996

Consider the following iterative process of removing vertices from Gex . Put G0 = Gex and n0 = n. Let v0 ∈ V (G) be an
arbitrarily chosen vertex of G0 satisfying d(v0 ) < (1 − 3k3−4 )n0 . Define G1 = G − v0 and n1 = n0 − 1. For j ≥ 1 if the
minimum degree in Gj satisfies δ (Gj ) ≥ (1 − 3k3−4 )nj then stop the process, otherwise take a vertex vj ∈ V (Gj ) of degree
dGj (vj ) < (1 − 3k3−4 )nj and define Gj+1 = Gj − vj and nj+1 = nj − 1.
We first show that this process must stop after at most n/2 steps. To see this note that the number of copies of Km
removed with each deleted vertex is exactly the number of ( k−copies of Km−1 in its neighborhood. By Proposition 3.1 for any
(k − 1)-chromatic graph H ′ , ex(n, Km−1 , H ′ ) = (1 + o(1)) m−1 ( k−n 2 )m−1 . As Gex is H-free, the neighborhood of any vertex
2
)
should be (H − v )-free, where v ∈ V (H) is such that χ (H − v ) = k − 1.
Thus at step j (starting to count from j = 0), as dGj (vj ) < nj (1 − 3k3−4 ), at most
3
k − 2 ( nj (1 − )
( ) )m−1
ex(dGj (vj ), Km−1 , H − v ) = (1 + o(1))
3k−4
m−1 k−2
copies of Km have been removed.
As the following equality holds
( )
1 3 1 1
− 1− =
k−1 3k − 4 k−2 (3k − 4)(k − 2)(k − 1)
(( )m−1
one can choose δ = δ (k, m) > 0 so that 3 1 1
− δ ). Thus the number of copies of Km removed
)
1− 3k−4 k−2
= (k−1)m−1
(1
at step j is no more than:
( ) ( )
k−2 1 k−1 m
(1 + o(1))nm
j
−1
(1 − δ ) = (1 + o(1))nm
j
−1
(1 − δ ). (2)
m − 1 (k − 1)m−1 m (k − 1)m
Together with the fact that
n/2−1 (1
∑ 1 1 )
(n − r)m−1 ≤ (1 + o(1)) − nm
m m 2m
r =0
n
we conclude that the number of copies of Km removed during the first 2
steps is at most
n/2−1 ( )
∑ k−1 m
(1 + o(1)) nm
j
−1
(1 − δ )
m (k − 1)m
j=0
( )
k−1 1 ( 1 )
≤(1 + o(1))(1 − δ ) 1− nm .
m (k − 1)m 2m
(k−1)( 1 )m
By Proposition 3.1, the graph Gn/2 has at most ex( 12 n, Km , H) = (1 + o(1)) m
n
2 (k−1)
copies of Km , and hence the total
number of copies of Km in Gex is at most
( ) ( )
k−1 1 ( 1 ) k−1 1 1
(1 + o(1))n (1 − δ )
m
1− + (1 + o(1))n m
m (k − 1)m 2m m (k − 1)m 2m
( )
( 1 )) k − 1 1
=(1 + o(1))nm 1 − δ 1 − m .
(
2 m (k − 1)m

But if ϵ is small(enough this

) contradicts (1). Thus the process must stop after r + 1 ≤
n
2
steps.
As δ (Gr ) ≥ 1 − 3
3k−4
nr and H is edge critical, Corollary 2.3 implies that χ (Gr ) ≤ k − 1. Define V (Gr ) = Vr .
The k − 1 partite subgraph of G[Vr ] with the maximum possible number of copies of Km has at least as many copies of Km
as Gr . By Lemma 2.8, part 1 we can now add the vertices removed during the steps of the process starting from j = r − 1
until j = 0, keeping the resulting subgraph (k − 1)-partite, where with each such vertex we add at least
( )
k−1 m
(1 + o(1))nm
j
−1
(1 − c1 ϵ )
m (k − 1)m
copies of Km . Assuming that ϵ is small enough to ensure, say, c1 ϵ < δ/2 it follows that in each such step the number of added
copies of Km exceeds the number of copies removed in the corresponding removal step.
When all the vertices are back we obtain a k − 1 partite subgraph of G containing more)copies of Km than Gex . This subgraph
is H-free, contradicting the maximality of Gex . Thus the inequality δ (Gex ) ≥ 1 − 3k3−4 n must hold and the result follows
(
from Corollary 2.3. □
 N. Alon and C. Shikhelman / Discrete Mathematics 342 (2019) 988–996 993

4. Maximizing the number of blow-ups of cliques

To prove Theorem 1.3 we first need a good estimate on ex(n, Km (t), H) for H satisfying χ (H) = m + 1.

Proposition 4.1. For integers m and t and any fixed graph H such that χ (H) = m + 1,
)m
n/m
(
ex(n, Km (t), H) = (1 + o(1)) .
t
(n/m)m
Proof. To show that ex(n, Km (t), H) ≥ (1 + o(1)) t
it is enough to take the m-sided Turán graph (i.e. the m-partite graph
(n/m)m
with sides of nearly equal size). As χ (H) = m + 1 it is H-free and has (1 + o(1)) t copies of Km (t).
For the upper bound, in [4] it is shown that the graph which is Km+1 free and has the maximum possible number of copies
of Km (t) is a complete multipartite graph. It is not difficult to see that the Turán graph maximizes the number of copies of
Km (t) among these. Thus
)m
n/m
(
ex(n, Km (t), Km+1 ) = (1 + o(1)) .
t
Let H be a fixed graph with χ (H) = m + 1, and let G be an H-free graph on n vertices with the maximum number of copies
of Km (t). By Corollary 2.6 G can be made Km+1 -free by deleting o(n2 ) edges, and with them at most o(n2 )O(nmt −2 ) = o(nmt )
copies of Km (t). Let G′ be the graph obtained by removing those o(n2 ) edges.
As G′ is Km+1 -free we get

(1 + o(1))ex(n, Km (t), H) = (1 + o(1))N (G, Km (t))

)m
n/ m
(
= N (G′ , Km (t)) ≤ ex(n, Km (t), Km+1 ) = (1 + o(1))
t
(n/m)m
and so ex(n, Km (t), H) ≤ (1 + o(1)) t
as needed. □

The idea of the proof of Theorem 1.3 is similar to the one of Theorem 1.2 but some of the estimates are more involved.

Proof of Theorem 1.3. Let G be a graph with δ (G) > (1 − ϵ )n and let Gex ∈ G (Km (t), H). If δ (Gex ) ≥ (1 − 3m3−1 )n then by
Corollary 2.3, χ (Gex ) ≤ m, as H is edge critical and
( we are done.
Assume towards contradiction that δ (Gex ) < 1 − 3m3−1 n. Consider the following iterative process, similar to the one in
)

the proof of Theorem 1.2. Put G0 = Gex and n0 = n. At step j > 0 if Gj satisfies δ (Gj ) ≥ 1 − 3m3−1 nj then stop the process,
( )

otherwise take vj ∈ V (Gj ) of degree dGj (vj ) ≤ 1 − 3m3−1 nj and define Gj+1 = Gj − vj and nj+1 = nj − 1. We show that the
( )
process must stop after at most 2n steps.
To bound the number of Km (t) removed at each step we take care of two cases. If all of the vertices in the color class of vi
in Km (t) are non-neighbors of it in Gi call the copy sparse. Otherwise call it dense.
First we estimate the number of dense copies. Let Km+−1 (t) be the graph obtained by taking Km−1 (t) and adding to
it a vertex that is connected to all of the other vertices. The number of dense copies of Km (t) containing vi is at most
N (G[N(v )], Km−1 (t)) · ni .
+ t −2

As H is edge critical there is a vertex v ∈ V (H) such that χ (H − v ) = m, let H ′ = H − v . By Proposition 2.5 if H is contained
in a blow-up of T then ex(n, T , H) = o(nv (T ) ). As the neighborhood of vi must be H ′ -free and as H ′ is contained in a blow-up
of Km+−1 (t), it follows that N (G[N(v )], Km+−1 (t)) = o(|N(v )|t(m−1)+1 ). Thus the number of dense copies of Km (t) in G containing
vi is o(ntm
i
−1
).
As for the sparse copies, let A(vi ) be the number of sparse copies of Km (t) in Gi containing vi . Let NGc (vi ) = V (Gi ) \ (NGi (vi ) ∪
i
{vi }) and d = dGi (vi ), and let H ′ = H − v for v ∈ V (H) such that χ (H ′ ) = m. Using Proposition 4.1 we obtain the following
bound on the number of sparse copies of Km (t) containing vi
∑
A(vi ) ≤ ex |N(vi ) ∩ N(u1 ) ∩ . . . ∩ N(ut −1 )|, Km−1 (t), H ′
( )

u1 ,...,ut −1 ⊆NGc (vi )

i
( )
ni − d − 1
≤ ex(d, Km−1 (t), H ′ )
t −1
)m−1
d/(m − 1)
( )(
ni − d
=(1 + o(1))
t −1 t
)m−1
(ni − d)t −1 dt
(
≤(1 + o(1)) .
(t − 1)! (m − 1)t t !
994 N. Alon and C. Shikhelman / Discrete Mathematics 342 (2019) 988–996

To bound this quantity consider the following function f (d) = dt(m−1) (ni − d)t −1 . Note that f (d) is a polynomial in d,
f (d) > 0 for 0 < d < ni and f (ni ) = f (0) = 0. Furthermore
f ′ (d) =(dt(m−1) (ni − d)t −1 )′
=dt(m−1)−1 (ni − d)t −2 [t(m − 1)(ni − d) − (t − 1)d].
ni =: β and is positive in [0, β]. It follows that between
( t(m−1) ) ( t −1
)
Thus f ′ (d) = 0 for d = 0, d = ni , d = t(m−1)+(t −1) ni = 1 − tm −1
0 and n, f (d) obtains its global maximum at the single values 0 < β < n for which f ′ (β ) = 0, and it is increasing in [0, β].
In our case d < (1 − 3m3−1 )ni and as m > 2 it follows that 1 − tm t −1
−1
> 1 − 3m3−1 . We conclude that f (d) ≤ f ((1 − 3m3−1 )ni ).
Plugging this value in our previous estimate of A(vi ) it follows that
t −1 ( t )m−1
( 3m3−1 ni ) ((1 − 3m3−1 )ni )
A(vi ) ≤(1 + o(1))
(t − 1)! (m − 1)t t !
( 3 )t −1 ( 3 )t(m−1) 1
=(1 + o(1))nmti
−1
1− .
3m − 1 3m − 1 (t − 1)!(m − 1)t(m−1) (t !)m−1
)t −1 ( )t(m−1)
Next we bound 3m3−1 1 − 3m3−1 . As 3m3−1 = (3m− 1
+ m1 the following holds:
(
1)m
( 3 )t −1 ( 3 )t(m−1) ( 1 1 )t −1 ( m − 1 1 )t(m−1)
1− = + −
3m − 1 3m − 1 m (3m − 1)m m (3m − 1)m
1 ( m − 1 )t(m−1) ( 1 )t −1 ( 1 )t(m−1)
= t −1 1+ 1−
m m 3m − 1 (3m − 1)(m − 1)
1 ( m − 1 )t(m−1) ( 1 )t −1
−t /(3m−1)
≤ t −1 1+ e
m m 3m − 1
( )(t −1)
1 ( m − 1 )t(m−1) ( 1 )
= t −1 1+ e−1/(3m−1) e−1/(3m−1)
m m 3m − 1
1 ( m − 1 )t(m−1)
≤ t −1 (1 − δ )
m m
for an appropriate δ := δ (m, t) > 0. Indeed such a δ exists as e−1/(3m−1) < 1 and 1 + 3m1−1 e−1/(3m−1) < 1 for m > 2.
( )

Therefore, the number of copies of Km (t) (both dense and sparse) removed at step i is at most
1 ( m − 1 )t(m−1) 1
(1 + o(1))nmt
i
−1
(1 − δ ) + o(nimt −1 )
mt −1 m (t − 1)!(m − 1)t(m−1) (t !)m−1
mt
= (1 + o(1))nmt
i
−1
(1 − δ ) . (3)
mtm (t !)m
n
∑n/2 mt (
− r)mt −1 ≤ (1 + o(1)) nmt 1 − 1
)
If the process continues for 2
steps, as r =0 (n 2mt
the total number of copies of Km (t)
removed is at most
(n/2)−1
∑ mt
(1 + o(1))(n − r)mt −1 (1 − δ )
mtm (t !)m
r =0
( 1 ) 1
≤(1 + o(1))nmt (1 − δ ) 1 − mt .
2 mtm (t !)m
By Proposition 4.1 in the graph Gn/2 the number of copies of Km (t) is at most
)m )m
n/(2m) nt
( (
1 1
(1 + o(1)) ≤ (1 + o(1)) = (1 + o(1))ntm .
t (2m)t t ! mtm (t !)m 2mt
Thus the number of copies of Km (t) in Gex is at most
(( )
1 )) 1
ntm 1−δ 1−
(
2mt mtm (t !)m
in contradiction to the maximality of Gex . So the process must stop after n/2 steps.
Assume that we have stopped at step r < n/2 and let Vr = V (Gr ). By Corollary 2.3 Gr must be m-partite, thus the m-partite
subgraph of G[Vr ] with the maximum possible number of copies of Km (t) has at least as many copies of Km (t) as Gr .
By Lemma 2.8, part 2, we can return the vertices removed in the process in a reverse order (starting from vr −1 until v0 )
keeping the graph m-partite and adding with each vertex vj at least (1 + o(1))nmt
j
−1 mt
(t !mt )m
(1 − c2 ϵ ) copies of Km (t). Assuming
that ϵ is small enough to ensure, say, c2 ϵ < δ/2 it follows that with each vertex vj we add more copies of Km (t) than were
removed at the corresponding step.
 N. Alon and C. Shikhelman / Discrete Mathematics 342 (2019) 988–996 995

Thus when all vertices are returned we obtain an m-partite graph with more copies of K)m (t) than Gex . As an m-partite
graph is H-free this contradicts the definition of Gex . Thus it must be that δ (Gex ) ≥ 1 − 3m3−1 n and Gex is m-partite. □
(

5. Forbidding graphs that are not edge critical

The proofs of Theorems 1.2 and 1.3 actually give a slightly stronger result than stated, as follows

Lemma 5.1. Let m < k and t be integers, let G be a graph on n vertices such that δ (G) > (1 − ϵ )n, where ϵ := ϵ (k, m, t) > 0
is sufficiently small, and let Gex ∈ Gex (Km (t), Kk ). Assume that k = m + 1 or t = 1. Then for every Kk -free subgraph of G on the
same set of vertices, say G1 ⊆ G, at least one of the following holds:

1. N (G1 , Km (t)) ≤ γ N (Gex , Km (t)) for some γ := γ (k, m, t) < 1.

2. G1 can be made (k − 1)-chromatic by deleting o(n2 ) edges.

Proof. If δ (G1 ) ≥ 1 − 3k3−4 v (G1 ) then by Theorem 2.1 G1 is (k − 1)-chromatic and hence case (2) holds and we are done. If
( )

δ (G1 ) < 1 − 3k3−4 v (G1 ) we consider a similar process to the one in the proofs of Theorems 1.2 and 1.3. For step j = 0 of the
( )

process define G10 = G1 , for steps j > 0 let vj be a vertex of minimum degree in G1j−1 and define G1j = G1j−1 − vj . The process
stops when either δ (G1j ) ≥ 1 − 3k3−4 v (G1j ) or when v (G1j ) = α n for α := α (γ ) small enough.
( )

The calculations)in the proofs of Theorems 1.2 and 1.3 (see Eqs. (2) and (3)) yield that when removing a vertex of degree
less than 1 − 3k3−4 v (G1j ) the number of copies of Km (t) removed with it is at most
(
( )
k−1 mt
(1 + o(1))(1 − δ )v (G1j )mt −1 .
m (k − 1)m (t !mt −1 )m
Assume that the process stops at step r. If r = o(n) then by Theorem 2.1 the graph G1r is (k − 1)-chromatic. In these r steps
o(n) vertices were deleted and with them no more than o(n2 ) edges, thus case (2) holds.
If r = cn for some c ≤ 1 −α define the graph Gr as follows. If c = 1 −α take Gr to be a (k − 1)-chromatic subgraph of G1 on
the vertices of G1r with the maximum possible number of copies of Km (t). It must be that N (Gr , Km (t)) − N (G1r , Km (t)) < α ′ nmt
for an appropriate α ′ = α ′ (α ) which tends to 0 as α tends to 0. If c > α take Gr = G1r , by Theorem 2.1 this graph is
(k − 1)-chromatic.
As Gr is (k − 1)-chromatic in both cases we can apply Lemma 2.8 to it and add back the vertices removed in the process,
starting from j = r − 1 to j = 1, while keeping the graph (k − 1)-chromatic. We get that the number of copies of Km (t) added
with each vertex vj is at least
( )
k−1 mt
(1 + o(1))(1 − c ϵ )v (G1j )mt −1 .
m (k − 1)m (t !mt −1 )m
Let G2 be the graph obtained after adding back all the vertices.
Assume that ϵ is small enough to ensure that δ − c ϵ > c ′ > 0 for some c ′ := c ′ (γ ). Let nj = v (G1j ) = n − j, and note that
∑r ∑cn
j=0nmt
j
−1
= j=0 (n
1
− j)mt −1 ≥ (1 + o(1))nmt mt (1 − (1 − c)mt ). Thus the difference in the number of copies of Km (t) in G1
2
and G is at least
r ( )
∑ k−1 mt
N (G2 , Km (t)) − N (G1 , Km (t)) ≥(1 + o(1)) c ′ nmt − α ′ nmt
−1
j
m (k − 1)m (t !mt −1 )m
j=0
( )
k−1 1
≥(1 + o(1))(c ′ − α ′ )(1 − (1 − c)mt )nmt
m (k − 1)m (t !mt −1 )m
=(1 + o(1))(1 − γ )N (Gex , Km (t))
where c and α are chosen so that the last equality holds.
′ ′

As G2 is a Kk -free subgraph of G, N (Gex , Km (t)) ≥ N (G2 , Km (t)) and thus γ N (Gex , Km (t)) ≥ N (G1 , Km (t)) and case (1) holds,
as needed. □
The proof of Proposition 1.4 is now a simple corollary of Lemma 5.1.

Proof of Proposition 1.4. Let G be a graph on n vertices with δ (G) > (1 − ϵ )n and let Gex(H) ∈ Gex (Km (t), H). By Lemma 2.7
N (Gex(H) , Km (t)) = Θ (nmt ). By Corollary 2.6 there is a graph G1 ⊆ Gex(H) which is Kk -free and e(Gex(H) ) − e(G1 ) = o(n2 ), and
thus
N (Gex(H) , Km (t)) = (1 + o(1))N (G1 , Km (t)).

Let Gex(Kk ) ∈ G (Km (t), Kk ). To apply Lemma 5.1 we show that

N (G1 , Km (t)) = (1 + o(1))N (Gex(Kk ) , Km (t)). (4)

996 N. Alon and C. Shikhelman / Discrete Mathematics 342 (2019) 988–996

By Theorems 1.2 and 1.3 Gex(Kk ) is (k − 1)-chromatic, and so it is H-free. Together with the fact that G1 is Kk -free, we get
N (G1 , Km (t)) ≤ N (Gex(Kk ) , Km (t)) ≤ N (Gex(H) , Km (t)) = (1 + o(1))N (G1 , Km (t))

implying (4).
Thus case (1) in Lemma 5.1 does not hold for G1 , and so case (2) must hold, i.e. G1 can be made (k − 1)-chromatic by
deleting o(n2 ) edges. As we got G1 from Gex(H) by deleting o(n2 ) edges we get the required result. □

6. Concluding remarks and open problems

• Corollary 2.3 and Theorems 1.2 and 1.3 cannot be directly generalized for graphs H which are not edge critical. In [5]
the following is shown (a weaker version of this statement is proved in [3])

Theorem ( 6.1 ([5]). Let H )be a fixed graph on h vertices such that χ (H) = k ≥ 3 and let G be an H-free
2/3
graph on n vertices with
δ (G) ≥ 1 − 3k3−4 + o(1) n, where n is large enough. Then one can delete at most O(n2−1/(4(k−1) h) ) edges from G and make it
(k − 1)-colorable.
This suggests that a stronger version of Proposition 1.4, stating that any extremal graph Gex as in the proposition can be
made (k − 1)-chromatic by deleting O(n2−µ(H) ) edges for some µ(H) > 0, is likely to be true.

• Theorem 1.3 is limited to the case where χ (H) = m + 1, and from this we also get the condition in Proposition 1.4. One
of the problems in extending it to graphs H with higher chromatic number is that of finding an explicit tight bound on
ex(n, Km (t), Kk ) for k > m + 1. In [4] it is shown that the extremal graph is (k − 1)-partite. However, it is not difficult
to check that for k ≥ m + 2 such that m ∤ k − 1 and large values of t, the parts are not of equal sizes.
• Theorems in the same spirit as those proven here may hold for other pairs of graphs T and H. In [4] it is observed that
if H is not contained in any blow-up of T then ex(n, T , H) = Θ (nv (T ) ). This of course does not mean that the extremal
graph is a blow-up of T , but in cases it is a similar behavior to that in the results proven here might be expected.
A notable example is the case T = C5 and H = K3 . In [16] and independently [15] it is shown that when 5|n the
extremal graph is the equal sided blow-up of C5 . It might be true that this behavior holds for subgraphs of graphs of
high minimum degree and not only for subgraphs of Kn , that is, the extremal subgraphs in this case may be subgraphs
of the equal sided blow-up of C5 .
• The problem of obtaining the best possible bounds for the minimum degree ensuring that the results stated in
Theorems 1.2 and 1.3 and Proposition 1.4 hold is also interesting, but appears to be difficult. Even the very special
case of Theorem 1.2 with H = K3 and m = 2, conjectured in [7] to be 3/4 + o(1), is open.

References

[1] N. Alon, Problems and results in extremal combinatorics, II, Discrete Math. 308 (2008) 4460–4472.
[2] N. Alon, E. Fischer, M. Krivelevich, M. Szegedy, Efficient testing of large graphs, Combinatorica 20 (4) (2000) 451–476.
[3] N. Alon, A. Shapira, B. Sudakov, Additive approximation for edge-deletion problems, Ann. of Math. (2009) 371–411.
[4] N. Alon, C. Shikhelman, Many T copies in H-free graph, J. Combin. Theory Ser. B 121 (2016) 146–172.
[5] N. Alon, B. Sudakov, H-free graphs of large minimum degree, Electron. J. Combin. (2006).
[6] B. Andrásfai, P. Erdős, V.T. Sós, On the connection between chromatic number, maximal clique and minimal degree of a graph, Discrete Math. 8 (3)
(1974) 205–218.
[7] J. Balogh, P. Keevash, B. Sudakov, On the minimal degree implying equality of the largest triangle-free and bipartite subgraphs, J. Combin. Theory Ser.
B 96 (6) (2006) 919–932.
[8] A. Bondy, J. Shen, S. Thomassé, C. Thomassen, Density conditions for triangles in multipartite graphs, Combinatorica 26 (2) (2006) 121–131.
[9] D. Conlon, J. Fox, Graph removal lemmas, Surv. Combinatorics 1 (2) (2013) 3–50.
[10] P. Erdős, On some problems in graph theory, combinatorial analysis and combinatorial number theory, in: Graph Theory and Combinatorics, Academic
Press, London, 1984, pp. 1–17, (Cambridge, 1983).
[11] P. Erdős, M. Simonovits, A limit theorem in graph theory, Studia Sci. Math. Hungar. 1 (1966).
[12] P. Erdős, M. Simonovits, On a valence problem in extremal graph theory, Discrete Math. 5 (4) (1973) 323–334.
[13] P. Erdős, A. Stone, On the structure of linear graphs, Bull. Amer. Math. Soc. 52 (1946) 1087–1091.
[14] J. Fox, A new proof of the graph removal lemma, Ann. of Math. 174 (2011) 561–579.
[15] A. Grzesik, On the maximum number of five-cycles in a triangle-free graph, J. Combin. Theory Ser. B 102 (5) (2012) 1061–1066.
[16] H. Hatami, J. Hladký, D. Král’, S. Norine, A. Razborov, On the number of pentagons in triangle-free graphs, J. Combin. Theory Ser. A 120 (3) (2013)
722–732.
[17] I.Z. Ruzsa, E. Szemerédi, Triple systems with no six points carrying three triangles, in: Combinatorics, in: Coll. Math. Soc. J. Bolyai, vol. 18, Volume II,
1976, pp. 939–945.
[18] P. Turán, On an extremal problem in graph theory, Mat. Fiz. Lapok 48 (1941) 436–452.